How the past impacts the future: modelling the performance of evolutionarily distinct mammals through time

Abstract

How does past evolutionary performance impact future evolutionary performance? This is an important question not just for macroevolutionary biologists who wish to chart the phenomena that describe deep-time changes in biodiversity but also for conservation biologists, as evolutionarily distinct species—which may be deemed ‘low-performing’ in our current era—are increasingly the focus of conservation efforts. Contrasting hypotheses exist to account for the history and future of evolutionarily distinct species: on the one hand, they may be relicts of large radiations, potentially ‘doomed’ to extinction; or they may be slow-evolving, ‘living fossils’, likely neither to speciate nor go extinct; or they may be seeds of future radiations. Here, we attempt to test these hypotheses in Mammalia by combining a molecular phylogenetic supertree with fossil record occurrences and measuring change in evolutionary distinctness (ED) at different time slices. With these time slices, we modelled future ED as a function of past ED. We find that past evolutionary performance does indeed have an impact on future evolutionary performance: the most evolutionarily isolated clades tend to become more evolutionarily distinct with time, indicating that low-performing clades tend to remain low-performing throughout their evolutionary history.

This article is part of a discussion meeting issue ‘The past is a foreign country: how much can the fossil record actually inform conservation?’

1. Introduction

Evolutionary distinctness (ED; [1]) is a measure of the isolation of a species in a phylogenetic tree, expressed in millions of years. Many conservation biologists are interested in the conservation of species that are highly evolutionarily distinct as these species represent a disproportionately large amount of evolutionary history [1,2]; they may be likened to ‘living fossils’ [3]. Furthermore, ED has often been correlated (albeit contentiously; [4]) with trait distinctiveness. Consequently, it is often argued (and disputed [5]) that by targeting many evolutionarily distinct species, we would probably be capturing greater trait diversity, and would then be preserving ecosystem functions and services [6]. However, interpretation of the ‘conservation value’ of distinct species is not straightforward, and targeting them for conservation attention may be legitimately questioned if they are deemed more likely to become extinct regardless of human activity [4,7]. Being members of species-poor lineages, we might consider the evolutionarily distinct as potentially ‘doomed’ to extinction if they are simply the tail-ends of once-diverse clades [8,9], and/or evolutionary ‘dead-ends’ [10]. Alternatively, the distinctness of such species may garner them with unique adaptations that would allow them to respond differently to future environmental or ecological change [1,11,12], allowing them novel opportunities to radiate into recently vacated ecological space [13]. In this scenario, ED may be related to G. G. Simpson's ‘phylogenetic fuse’: the idea that lineages may persist for long periods of time ‘tinkering’ at low species number before exploding in diversity at a later time point [14]. Yet another interpretation is that evolutionarily distinct species may be members of long-lasting lineages that neither go extinct nor diversify in great numbers (‘panchronic’ sensu [15]). Finally, a neutral model could be proposed in which the origin and future potential of evolutionarily distinct species is not influenced by any overarching processes or phenomena.

Little is known about the future evolutionary potential of evolutionarily distinct lineages. Multiple studies have found that lineage age does not correlate with extinction risk [16–18]. Additionally, we know that extinction risk today is non-randomly distributed across the tree of life [19–22], although this does not appear to translate into higher extinction risks for evolutionarily distinct species [23–26]. However, these latter studies are mainly based on recent rates and patterns of extinction vulnerability and resilience, and species and clades at risk of extinction can radically vary over longer periods of geological time, such as during mass extinctions [27]. In particular, although the fossil record represents a uniquely important long-term archive that can be used to trace patterns of lineage diversity and diversification through time, few studies have hitherto explicitly attempted to test how past evolutionary performance has impacted future evolutionary performance using the fossil record, particularly from a conservation perspective (see [28] for one such recent example).

ED is most often calculated at the species-level [1], but we can also define a clade's ED as the average ED of all descendant species. It is therefore possible to assess the relationship between the ED of a clade at a given time point (ED_t0) and its subsequent ED at a later time point (ED_t1). We might expect a linear relationship between the ED values of the two time points if ED does not influence diversification (dotted line, figure 1). Such a model, however, does not take into account differences in the ages of trees between different time points and the number of species in a tree. A better null model is able to take into account these differences and may not be linear. In figure 1, we outline a series of possible scenarios for the relationship between past and future ED based on simulated birth–death tree models. The null model is generated from simulations in which ED has no impact on future speciation or extinction rates (‘null’, figure 1). If future evolutionary potential of species is influenced by how evolutionarily isolated a clade is, the relationship should diverge from this line. For example, under a strict ‘panchronic’ scenario, clades with low ED are more likely to contain species that will go on to speciate more in the future, so should have lower mean ED at the next time point as they will then share a similar proportion of the tree among more species, whereas clades with high ED are less likely to speciate and should hence show an increase in their mean ED over time (‘pan.’, figure 1). Alternatively, under a strict ‘evolutionary relict’ scenario, species with mid-ranging ED become increasingly more evolutionarily distinct up to the point of extinction and, therefore, show ED that is increasingly higher than expected in successive time points, while high ED should tend towards expectation as the whole clade goes extinct (‘rel.’, figure 1). Finally, we might consider the less discussed idea that highly evolutionarily distinct clades are more likely to diversify owing to their differences from other taxa, ‘phylogenetic fuse’ [14] (‘PF’, figure 1).

Figure 1.

Polynomial, linear modelling of possible general relationships between ED at time points in the past (t0, x-axis) and in the future (t1, y-axis) as determined from simulated birth–death trees under different scenarios. ED is natural logged to create a normal distribution. Trees were rescaled to make their branch lengths comparable to the Mammalian tree of life. Dark-grey indicates 95% confidence interval. Dotted lines indicate past ED matching future ED and are added for reference. ‘Null’ indicates the model where speciation and extinction rates are independent from ED. This model is repeated in subsequent panels as the solid, black line. Tendencies above this line indicate higher than expected ED at the next time step, and below the line, lower than expected ED at the next time step. In the panchronic scenario (pan.), mid-to-high values lead to higher values. In the evolutionary relict scenario (rel.), mid-to-high ED leads to increasing ED up to a maximum value at which point the whole clade becomes extinct. In the phylogenetic fuse scenario (PF), low-to-mid ED leads to higher values, and high values lead to low values. These curves were generated from birth–death tree-growth simulations where biases can be introduced based on tip ED. Scenarios are based on those described in [3]. For more details on how these simulations were created, see the electronic supplementary material, methods and results. (Online version in colour.)

Here, we compare ED between mammalian clades at different time points to investigate how ED changes at the clade level through time. We use mammals owing to their well-established phylogenetic tree [29] and the availability of relatively well-sampled fossil data [30]. We generate large molecular-fossil phylogenetic trees by taking a large mammalian supertree and adding fossil tips using a taxonomically constrained stochastic process. We then use these large trees to take time slices at comparable time points, and track the EDs of identifiable clades and species across these time points to generate an ED_t0 and ED_t1 dataset. With this dataset, we then test whether the ED_t1 ∼ ED_t0 relationship is nonlinear and whether it corresponds to any previously identified scenarios of evolutionary diversification in relation to ED as outlined above. Our results establish a robust new baseline for understanding the likely potential of evolutionarily distinct species today in terms of their future contribution to global biodiversity and its implications on their conservation.

2. Methods and materials

All of our analyses were performed in the R environment (v. 3.5) and the scripts for reproducing these results can be found on the primary author's GitHub page (https://github.com/DomBennett/Project-karenina).

Mammalian fossil data were downloaded from the Paleobiological Database (PBDB) [31] using the R package paleobioDB [32] on 25 September 2018. All records were constrained to ‘Mammalia’. In total, 109 536 fossil occurrence records were downloaded. These records were converted to species temporal records by merging records with shared species names and identifying minimum and maximum species ages. Taxonomic named ranks were determined for every species record, and records lacking taxonomic information below the family level were removed (438 records). Records with temporal ranges extending beyond the age of the mammalian supertree (166.2 Myr) were also excluded (121 records). In total, 19 028 species records remained, representing 128 families and 6659 genera. The temporal distribution was skewed towards the Recent; most observations occurred in the first 18 Myr (0.0059 Myr – 0%, 7.0 Myr – 25%, 17.7 Myr – 50%, 39.7 Myr – 75%, 164.8 Myr – 100%).

(a). Phylogenetic placement of fossils

Fossil species were added to a time-calibrated, molecular-based phylogenetic supertree of 4510 extant mammal species [29]. Taxonomic information was added to the tree by identifying the most likely taxonomic group for every node, by matching descendants to named entries in the National Center for Biotechnology Information (NCBI) taxonomic database [33] via the Global Names Resolver [34]. For every node, the lowest shared taxonomic group was selected from all taxonomic named ranks of matching named entries. In the NCBI taxonomy under mammals, there is up to 12 taxonomic ranks (from superorder to subgenus) for each species. Taxonomic groups were identified for 6603 of the 6618 nodes in the tree (greater than 99%).

Tips representing fossil species were added to the supertree using the pinTips() function in the R package treeman [35]. The position for each fossil species was constrained to edges directly parent to or descending from the node(s) with the lowest-ranked matching taxonomic label (with any edge or part of an edge younger than the minimum age of the fossil species excluded). Where these constraints left multiple possible attachment points, one was chosen at random. For example, if a fossil is matched to a given genus name, then the fossil tip could be added to the edge leading to this genus or to any of the edges that make up the genus. The time of extinction for fossil species was determined as a random point between its split from the tree and its youngest possible age (electronic supplementary material, figures S1–S2). Fossil species added to the tree thus all became extinct within their estimated age range, and their origination occurred before or during their estimated age range.

The stochastic fossil-adding process was iterated to generate a distribution of possible molecular-fossil trees; 100 iterations were deemed sufficient as initial analyses demonstrated that trees generated through this process had similar distributions of ED values. In total, a mean 18 964 ± 16 fossil tips were added to the original 4510 tipped mammalian supertree for each iteration. To assess whether the stochastic fossil-adding process was biasing results, the process was repeated with fossil lineages and age ranges randomly assorted, so that fossil placement was not constrained by their actual ages or taxonomy. Before analysing changes in ED between time points, we first tested whether the random set of trees differed significantly from the non-random set (electronic supplementary material, methods and results). If the random set were to not differ, then this would indicate that the fossil record informed pinning is no different from random placement.

(b). Determining change in evolutionary distinctness

For each tree generated through fossil-addition, a ‘time slice’ was taken to generate an ultrametric tree representing a phylogeny up to a given time point (electronic supplementary material, figure S2). For each of these slices, EDs were calculated for all tips using the method of [1], with ED of internal nodes calculated as the mean ED of all descendants of that node. In order to generate ultrametric trees that were as equally sampled as possible, time slices were taken at the midpoints of each epoch (Upper Jurassic to Recent; electronic supplementary material, table S1). ED_t0 and ED_t1 datasets were generated from the two distributions by matching ED values from one epoch to the next for all species/clades that spanned two or more epochs. Mean ED values were calculated for species/clades that occurred more than once across the tree distributions (later termed shared nodes). All values were converted to their natural logarithm before analysis, to convert the skewed distribution of ED to a normal distribution. Our datasets of species/clades were supplemented with taxonomic information (genus and order), the number of species present in the tree at t0, and the time (in Myr) between t0 and t1.

(c). Modelling

(i). Mixed-effects models

ED values of epoch-to-epoch transitions are not straightforward to compare: the ED values in one epoch will not be independent of those from another because many of the taxa are the same; transitions differ in tree age and starting number of tips at t0; the number of fossil records differ between each epoch (older have fewer than recent) and transition time-steps are not the same (electronic supplementary, material, table S1). To account for these factors and control for differences between epochs, we used linear mixed-effects models (LMEMs) [36,37]. We modelled future ED as a function of past ED (ED_t1 ∼ ED_t0) with the option of including a random effects structure that was able to incorporate both epoch and taxonomic information. We used an iterative model approach, starting with a basic LMEM model and adding additional terms to the random effect structures until there was no longer any significant or notable gain in explained variance. We compared models using ANOVA and the Akaike information criterion (AIC) [38]. For more details on LMEMs and our approach, see the electronic supplementary information, methods and results.

(ii). The three models

Using the LMEM framework, we generated two models to investigate how past ED affects future ED: a linear observed model (obs1), a nonlinear observed model (obs2) and an expected model (exp). We generated obs1 using the modelling framework as described above to generate the best-fitting linear model of ED_t1 ∼ ED_t0 given different random effects structures. We then took the same random effect structure to generate obs2 by adding increasing numbers of orthogonal polynomial degrees to explore increasingly complex nonlinear relationships, until there was no longer any increase in explained variance.

An expected model, akin to the ‘null’ birth–death model in figure 1, was built for comparison, to represent ED_t1 ∼ ED_t0 in a ‘neutral’ scenario where past average ED for a clade has no impact on future ED other than simply determining the starting points from which future ED can diverge. This model was a function of ED_t1 and the factors that influence ED between epochs: the number of tips in the tree at t0 (n), the time difference between epochs (tm) and an ED_t0 dummy variable (dummy). The dummy variable was a coarsened representation of ED_t0 calculated by rounding ED_t0 to the nearest integer and dividing the values by the maximum to generate values between 0 and 1. The variable provided limited categorical information on ED_t0 to the model (e.g. high ED, mid ED and low ED) and acted as a basis from which ED_t1 can then be estimated in conjunction with the other factors.

With these models, we first tested whether ED_t1 ∼ ED_t0 is nonlinear by comparing the goodness of fit to the data of obs1 and obs2, and then tested which of the simulated scenarios (figure 1) best describe ED_t1 ∼ ED_t0 by visually comparing the best observed (obs1 or obs2) to exp.

3. Results

(a). Estimating the linear model

The generated dataset consisted of 115 810 species/clade ED values, recorded across the nine epoch-to-epoch transitions. The distribution of shared nodes across the iterated trees was bimodal, with the majority of species/clades occurring 12 or fewer times (1–0%, 3–25%, 12–50%, 39–75%, 100–100%). There was a positive nonlinear relationship between ED_t0 and ED_t1 and differences between the epoch-to-epoch transitions were substantial, particularly at low ED values. Two epoch-to-epoch transitions (Jurassic Upper to Cretaceous Lower (JU–CL), Cretaceous Lower to Cretaceous Upper (CL–CU)) showed a different relationship from the others (electronic supplementary material, figure S3), probably owing to limited availability of data points for these epochs (6321 and 2288 for CL–CU and JU–CL, respectively, versus a mean of 15 314 for all other epoch transitions) and the much longer time separating them (39.5 Myr and 31.5 Myr for CL–CU and JU–CL, respectively, versus a mean of 11.9 Myr for all other transitions). We consequently removed JU–CL and CL–CU transitions from all subsequent analyses.

We determined the best linear model of ED_t1 ∼ ED_t0 to be m2i (table 1). This model incorporated a random effects structure that consisted of independent slopes for both epochs and a hierarchical taxonomic random effect (order/genus). However, we did not consider model choice to be crucial in interpretation of results, as estimated slopes were similar across all models (0.63–0.73).

Table 1.

Increasingly complex models for estimating the best-observed linear model, obs1. (Formulae (f) [37], intercepts (int), slopes (slp) degrees of freedom (d.f.), Akaike information criterion (AIC) and significance (p) are indicated. Significance is indicated by comparing the current row's model to the last significant model. The selected linear model (m2i) is highlighted in bold. ***p < 0.001, **p < 0.01, *p < 0.05.)

id.	f	int	slp	d.f.	AIC	p
m0	EDt1 ∼ EDt0	1.1068	0.7342	1	49 418
m1a	EDt1 ∼ EDt0 + (1|epoch)	1.2101	0.7148	4	21 337
m1b	EDt1 ∼ EDt0 + (EDt0|epoch)	1.2274	0.7061	6	17 492	***
m2a	EDt1 ∼ EDt0 + (EDt0|epoch) + (1|order)	1.2219	0.7100	7	16 568	***
m2b	EDt1 ∼ EDt0 + (EDt0|epoch) + (1|genus)	1.3483	0.6642	7	10 459	***
m2c	EDt1 ∼ EDt0 + (EDt0|epoch) + (1|order/genus)	1.3562	0.6638	8	10 296	***
m2d	EDt1 ∼ EDt0 + (EDt0|epoch) + (1|id)	1.2646	0.6920	7	16 372
m2e	EDt1 ∼ EDt0 + (EDt0|epoch) + (1|order/id)	1.2626	0.6951	8	15 557
m2f	EDt1 ∼ EDt0 + (EDt0|epoch) + (EDt0|order)	1.2111	0.7144	9	16 378
m2g	EDt1 ∼ EDt0 + (EDt0|epoch) + (EDt0|genus)	1.4242	0.6274	9	7255	***
m2h	EDt1 ∼ EDt0 + (EDt0|epoch) + (EDt0|id)	1.2726	0.6882	9	16 379
m2i	EDt1 ∼ EDt0 + (EDt0|epoch) + (EDt0|order/genus)	1.4010	0.6405	12	7106	***

(i). Estimating the nonlinear model

We compared the best linear model (m2i) with a range of nonlinear models, using the same random effects structure, based on orthogonal polynomials generated from ED_t0 for different exponents. Using ANOVA and AIC values, as expected, we found all the polynomial models to have significantly better fits than the linear model, indicating that evolutionarily distinct clades do have differences in diversification potential (table 2). Of the nonlinear models, we determined m3b, the third-order polynomial model, to be the best; this model had one of the lowest AICs and a low χ² p-value indicating goodness of fit. While there were significant improvements in AIC in higher orders of polynomial model, the drops in AIC were small (less than 1%), and were not considered sufficient to warrant the greater model complexity. Additionally, we repeated the final plotting analysis (figure 2) and found similar trends for all higher-order polynomial (greater than 3) models, indicating that any conclusions are not affected by our choice of the lower-order polynomial model.

Table 2.

Increasingly complex models for estimating the best-observed nonlinear model, obs2. (Order of polynomial (poly), degrees of freedom (d.f.), Akaike information criterion (AIC) and significance (p) are indicated. All models share the same model formula as m2i. Significance is indicated by comparing the current row's model to the last significant model. The selected nonlinear model (m3b) is highlighted in bold. ***p < 0.001, **p < 0.01, *p < 0.05.)

id.	poly	d.f.	AIC	p
m2i	1	12	7106
m3a	2	13	4727	***
m3b	3	14	4513	***
m3c	4	15	4501	***
m3d	5	16	4493	**
m3e	6	17	4473	***
m3f	7	15	4674

Figure 2.

Predicted log(ED_t1) values generated from the best-observed linear (obs1, solid red line) and expected model (i.e. ‘dummy variable’, n2b, solid black line) for a representative dataset of a range of log(ED_t0) values, a random subset of 100 genera, and all epoch-to-epoch transitions. When compared to figure 1, the observed relationship most closely resembles the ‘panchronic’ scenario. Dotted line indicates log(ED_t0) = log(ED_t1). Estimates across the different genera and epoch-to-epoch transitions are median averaged and 95% confidence intervals were calculated, with values representing natural logged millions of years (1 = 2.7, 2 = 7.4, 3 = 20.1, 4 = 54.6, 5 = 148.4). (Online version in colour.)

(ii). Estimating the expected model

We found the best-expected model to be n2b, which incorporates the ED_t0 dummy variable (dummy) with three orders of polynomial, difference in time between epochs (tm) and the starting number of species in a Mammalian tree at the beginning of an epoch (n), as well as a genus random effects structure of random slopes for dummy (table 3). Although it is likely that a model with random slopes for all three fixed effects would have produced a better fit, this model and models with multiple random effects did not converge indicating model over-fitting. Again, we found only marginal improvement with the use of a hierarchical taxonomic random effects structure (order/genus). Consequently, we limited our random effects structure to one non-hierarchal factor.

Table 3.

Expected models, exp, of ED_t1 against difference of time between epochs (tm), number of species in the tree at t0 (n), and an ED_t0 dummy variable generated from the rounded figures of ED_t0. (Formulae (f) [37], degrees of freedom (d.f.), Akaike information criterion (AIC) and significance (p) are indicated. Significance is indicated by comparing the current row's model to the last significant model. The ‘—’ indicates the separation of linear models and linear mixed-effects models. The selected expected model (n2b) is highlighted in bold. ***p < 0.001, **p < 0.01, *p < 0.05.)

id.	f	d.f.	AIC	p
n0a	EDt1 ∼ tm	3	118 248
n0b	EDt1 ∼ n	3	114 116
n0c	EDt1 ∼ tm + n	4	113 934	***
n0d	EDt1 ∼ dummy	3	71 796
n0e	EDt1 ∼ dummy + tm + n	5	49 229	***
n1a	EDt1 ∼ dummy + tm + n + (1|order)	6	48 182	—
n1b	EDt1 ∼ dummy + tm + n + (1|genus)	6	35 783	***
n1c	EDt1 ∼ dummy + tm + n + (1|id)	6	47 644
n1d	EDt1 ∼ dummy + tm + n + (dummy|genus)	8	31 649	***
n1e	EDt1 ∼ dummy + tm + n + (tm|genus)	8	32 141
n1f	EDt1 ∼ dummy + tm + n + (n|genus)	8	31 262
n2a	EDt1 ∼ poly(dummy 2)+tm + n + (dummy|genus)	9	30 721	***
n2b	EDt1 ∼ poly(dummy 3)+tm + n + (dummy|genus)	10	30 496	***
n2c	EDt1 ∼ poly(dummy 4)+tm + n + (dummy|genus)	11	30 472	***

(iii). Comparing the expected and the observed

We compared the expected model (exp, n2b) with the best-observed polynomial model (obs2, m3b) by plotting the parametrized lines to create figures analogous to figure 1. Because these models generate tens of thousands of individual lines (one line per genus per epoch transition), plotting all outputs directly would make visual interpretation of trends difficult, so we constructed a representative dataset consisting of 100 equally spaced ED_t0 values spanning from the observed minimum to observed maximum for 100 random selections of all genera across all epochs. This representative dataset consisted of 70 000 rows. We used the dataset to predict ED_t1 for obs2 and plotted the predicted values by calculating the median and 95% confidence intervals across genera and epochs (figure 2).

For both the best nonlinear model when compared to the expected model, we found that, on average, high averaged clade ED_t0 values (above ∼ e^2.5 or approx. 12 Myr) tended to lead to higher than expected ED_t1 values, a pattern that was most consistent with the ‘panchronic’ scenario; showing increasing divergence from the expected model at higher ED_t0 values. For low values of averaged clade ED_t0 (below ∼ e^1.5 or approx. 5 Myr), we found that ED_t1 was lower than expected, indicating clades with low ED tended to speciate more than clades with high ED. These patterns were consistent across the different epochs (electronic supplementary material, figure S6). Additionally, upon repeating the above analyses for non-taxonomically informed stochastic fossil placement in the molecular-fossil trees, we were able to rule out that our results were the product of random fossil placement (electronic supplementary material, methods and results).

4. Discussion

In order to determine the future evolutionary potential of clades composed of evolutionarily distinct species, we modelled future ED as a function of past ED using data generated from the mammalian fossil record. We find taxa that are or have been evolutionarily distinct are more likely to become more evolutionarily distinct in the future. These findings fit closest to a ‘panchronic’ scenario for the evolution of the evolutionarily distinct; evolutionarily distinct taxa in this scenario are disproportionately composed of ‘living fossils’ that experience both reduced rates of extinction and speciation [15]. Observed results differ from what would be expected from an ‘evolutionarily relict’ (dead-end) scenario, in which species most commonly become evolutionarily distinct by being the sole survivors of a once-large clade (figures 1 and 2). The observed model shows a slow gain in ED for clades that are already very evolutionarily distinct; however, in an ‘evolutionary relict’ scenario, clades with mid to low ED—below e² or 7 Myr—at t0 should tend to reach high ED at t1. Instead, these clades have lower ED at t1. Equally, our results are not compatible with a ‘phylogenetic fuse’ scenario in which evolutionarily distinct clades tend to be the seeds of future diversity, because taxa with high ED were not observed to lead to taxa with low ED over time.

(a). Potential biases

Our results indicate that the randomness of the fossil-adding process is likely to have reduced (not generated) the nonlinearity of the observed results. Firstly, upon comparing the models fitted to the taxonomically informed versus randomly added distribution of molecular-fossil trees, we showed that the pattern for high ED values in t0 leading to even higher ED values in t1 was stronger in the taxonomically informed distribution. Secondly, the observed nonlinearity was greatest for the data points in which we have the greatest confidence, as high ED values in t0 led to even higher ED values in t1 for the species/clades that were shared more than 50 times across tree iterations (electronic supplementary material, ‘Comparing model outcomes between the real and random’).

The level of detail of the fossil taxonomy limits the placement of fossils (electronic supplementary material, figure S1). This issue is more likely to affect the ED estimates of lower-level clades (e.g. at the genus or subfamily level) than higher-level clades, because although fossil species may be added to the wrong subfamily owing to lack of taxonomic resolution, they are very unlikely to be added to the wrong family. Nonetheless, the taxonomic misplacement of fossil species will impact estimates of ED. Another source of error in the placement of fossil species is their age of speciation and extinction. Ages of appearance are likely to be underestimated owing to the Signor–Lipps effect [39]. Although this effect is not directly accounted for using any statistical or mechanistic models, the effect's impact should be mitigated by our fossil-adding process, in which origination of fossil lineages was set to occur before or during the estimated age ranges. Both of these sources of error could theoretically introduce bias that might account for some or all of the nonlinearity observed in our model outputs, but it is difficult to conceive of a mechanism by which taxonomic or dating errors could produce the strong observed patterns. Additionally, we note that phylogenetic trees which are not wholly accurate can still produce metrics more similar to true values than would be expected [40].

Taphonomic (preservational) inconsistencies in the fossil record are potentially another source of bias [41]. To produce the observed nonlinearity in our model outputs, a bias where species with high ED were sampled less frequently than species with low ED would be required. One mechanism that could achieve a fossil bias of this kind might be differential sampling of evolutionarily distinct species if they have smaller body sizes and/or population sizes [42]. However, no studies have found significant correlations between ED and population size, and many studies suggest that evolutionarily distinct species tend to have greater body mass for a range of taxonomic groups [11,43–45], which may be associated with increased likelihood of representation in the fossil record, e.g. [46]. In mammals body mass is only weakly or not at all correlated with ED [12]. Additionally, taphonomic biases are likely to be strongly taxonomically controlled, for example, because related species are likely to occupy similar habitats [47]. However, the nonlinear relationship in our model outputs is detected even when using a random effects structure that controls for inter-relatedness. We therefore consider it very unlikely that this potential scenario for bias is responsible for generating the patterns that we observe.

(b). Implications for conservation biology

Palaeontology and macroevolution have an important potential role to play in informing conservation, and our study demonstrates how fossil data can be used to provide unique insights for conservation prioritization and decision-making. Funds are limited and not all species can be saved; prioritization is therefore key to conservation strategy [1,12,48]. To date, most species-level conservation prioritization has been focused on well-known charismatic taxa such as tigers, rhinos and polar bears [49,50], as well as on highly threatened species in urgent need of conservation [51]. One approach to move away from this bias has been to place greater conservation effort on evolutionarily distinct species (e.g. [52–54]), based in part upon the supposition that such taxa might contribute to future evolutionary potential; that is, they represent ‘cradles’ of future diversity, as opposed to ‘museums’ of past diversity [55]. One such widely publicized effort has been to rank species for conservation prioritization based on their Evolutionarily Distinct and Globally Endangered (‘EDGE’) score [1,12]. However, arguments arise over the usefulness of investing efforts into species that may already be evolutionarily ‘doomed’ [4,7]. Indeed a recent study found that conservation strategies which prioritize the preservation of evolutionarily distinct species led to a reduction in the number of future lineages [28].

The findings of our investigation of large-scale fossil data provide an important new perspective on this conservation question. It is clear that, at a broad global scale and based on data for mammals, focusing conservation attention on evolutionarily distinct species will not safeguard the overall future ‘evolvability’ of life on Earth, as implied by a ‘phylogenetic fuse’ scenario [14]. However, our results are equally not indicative of an evolutionary ‘dead-end’ scenario for the origin and future of evolutionarily distinct species over geological time [8,9]. These taxa are neither doomed nor the likely seeds of future radiations; our findings demonstrate that they are likely to remain evolutionarily distinct, and appear to be merely the slow-evolving ends of the tree of life. Although the exact mechanism for what may cause the increasing distinctness of the distinct is not tested here, the results corroborate with studies demonstrating the widespread phenomenon of age-dependant speciation [56–58].

Our findings may appear to argue against special prioritization of evolutionarily distinct species, contrary to the philosophy of the ‘EDGE’ approach. However, we definitely do not advocate withdrawing conservation attention from these species. While our study provides a new overview of the likely persistence and future diversification of evolutionarily distinct species and clades, thus challenging one of the general justifications that has been proposed for conserving such species, these findings provide a broad pattern across the Mammalia as a whole; we are unable to predict whether any specific species or clade within this major animal group may indeed diversify in the future, to what extent, or at what time, in response to any number of possible future ecological scenarios. Indeed, we note that whereas our analyses are based upon data for speciation and extinction events across deep time caused by non-anthropogenic processes, current-day global biodiversity loss is driven by anthropogenic processes and is likely to be associated with different patterns of species vulnerability or resilience [59,60], including the generation of ‘artificially evolutionarily distinct’ or ‘neo-relict’ clades which were species-rich before the Late Quaternary but have experienced disproportionate levels of human-caused extinction (e.g. proboscideans, sloths) [22].

Furthermore, future potential is only one criterion for targeting conservation effort, both quantifiable and intangible factors should also be considered. Conserving evolutionary history and the breadth of mammalian biodiversity is an important goal in and of itself, for a variety of biological reasons (e.g. trait diversity and its relationship with ecosystem function and complexity; [11]) and other reasons (e.g. ethical, financial and cultural), irrespective of the possible future evolutionary trajectory of a given species or clade. Evolutionarily distinct lineages have just as much of a right to survive the Anthropocene as do any others, and their unique representation of more independent evolutionary history than ‘normal’ lineages may still be considered worthy of particular attention.

Data accessibility

This article has no additional data.

Authors' contributions

D.J.B. devised and developed the project ideas, implemented the analyses and wrote the manuscript. M.D.S. and S.T.T. proposed the original project outline, contributed to the analyses and contributed to the manuscript.

Competing interests

We declare we have no competing interests.

Funding

This project was funded by a Natural and Environmental Research Council (NERC, UK) PhD grant.